Hyper-Ada: Active Data Augmentation via High-Order Feature Refinement with Adaptive Hypergraphs
Year: 2025
Type: article
Abstract: Data augmentation (DA) is a critical technique for addressing data scarcity in deep learning. However, traditional random augmentation methods are inefficient, while active learning (AL) often overlooks the high-order correlations among samples. To address these limitations, we propose the Progressive Active Data Augmentation (PADA) framework, which applies the intelligent selection principles of AL to the data augmentation process. Within this framework, we designed the core selection strategy, Hyper-Ada. This strategy leverages an Adaptive Hypergraph Convolution (AdaHGConv) network to refine new features embedded with high-order relations, derived from the global context of all augmented samples. We innovatively combine the model’s prediction "Certainty" with our proposed "Representation Shift"—the magnitude of feature change before and after refinement—as the selection criterion to identify high-quality "anchor" samples. Experiments on CIFAR-10, SVHN, and CIFAR-100 demonstrate that Hyper-Ada significantly outperforms supervised baselines and traditional AL methods, particularly in data-sparse scenarios. This validates the efficacy of guiding data augmentation through high-order feature refinement.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7125709171